NBER WORKING PAPER SERIES EARLY ... - Wellesley College

NBER WORKING PAPER SERIES

EARLY CHILDHOOD EDUCATION BY MOOC: LESSONS FROM SESAME STREET Melissa S. Kearney Phillip B. Levine Working Paper 21229 http://www.nber.org/papers/w21229 NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 June 2015

The authors thank Riley Wilson for exceptional research assistance throughout this project. Amy Wickett also provided valuable assistance creating our coverage measure. This paper also benefitted tremendously from discussions with Liz Cascio, Doug Miller, Ben Olken, Diane Schanzenbach, and Sergio Urzua as well as the comments of participants of seminars at MIT, Harvard University, University of Maryland, Wellesley College, Northwestern, the University of Illinois, the Federal Trade Commission, Georgetown University, and Williams College. We also thank Michael Henry at the University of Maryland Libraries (Special Collections in Mass Media & Culture) for his assistance in collecting archived Sesame Street information. We gratefully acknowledge financial support from the Spencer Foundation. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research. NBER working papers are circulated for discussion and comment purposes. They have not been peerreviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications. © 2015 by Melissa S. Kearney and Phillip B. Levine. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source.

Early Childhood Education by MOOC: Lessons from Sesame Street Melissa S. Kearney and Phillip B. Levine NBER Working Paper No. 21229 June 2015 JEL No. I24,J24 ABSTRACT Sesame Street is one of the largest early childhood interventions ever to take place. It was introduced in 1969 as an educational, early childhood program with the explicit goal of preparing preschool age children for school entry. Millions of children watched a typical episode in its early years. Well-designed studies at its inception provided evidence that watching the show generated an immediate and sizeable increase in test scores. In this paper we investigate whether the first cohorts of preschool children exposed to Sesame Street experienced improved outcomes subsequently. We implement an instrumental variables strategy exploiting limitations in television technology generated by distance to a broadcast tower and UHF versus VHF transmission to distinguish counties by Sesame Street reception quality. We relate this geographic variation to outcomes in Census data including grade-for-age status in 1980, educational attainment in 1990, and labor market outcomes in 2000. The results indicate that Sesame Street accomplished its goal of improving school readiness; preschool-aged children in areas with better reception when it was introduced were more likely to advance through school as appropriate for their age. This effect is particularly pronounced for boys and non-Hispanic, black children, as well as children living in economically disadvantaged areas. The evidence regarding the impact on ultimate educational attainment and labor market outcomes is inconclusive. Melissa S. Kearney Department of Economics University of Maryland 3105 Tydings Hall College Park, MD 20742 and NBER [email protected] Phillip B. Levine Department of Economics Wellesley College 106 Central Street Wellesley, MA 02481 and NBER [email protected]

E arly C hildhood E ducation by M O O C : L essons from Sesame Street Melissa S. Kearney and Phillip B. Levine I. I N T R O D U C T I O N In recent years, early childhood education, designed to improve subsequent life outcomes for students who participate, has received considerable attention. Programs like Perry Preschool, Head Start, universal pre-kindergarten, and others have taken center stage. Academic research has generally supported the role that early childhood education can play in improving outcomes for disadvantaged children, as reviewed by Duncan and Magnuson (2013), and that has led to specific proposals from those in the policy community (cf. Cascio and Schanzenbach, 2014). Both sides of the political spectrum have promoted its benefits (cf. Council of Economic Advisers, 2015; and Stevens, 2015). For all of this attention, it is surprising that perhaps the biggest, yet least costly, early childhood intervention, Sesame Street, has largely gone unnoticed. This show initially aired in 1969; its fundamental goal was to reduce the educational deficits experienced by disadvantaged youth based on differences in their preschool environment. It was a smash hit immediately upon its introduction, receiving tremendous critical acclaim and huge ratings. It cost pennies on the dollar relative to other early childhood interventions. Well-designed research studies conducted at that time, reviewed in detail below, indicate that the show had a substantial and immediate impact on test scores, comparable in size to those observed in early Head Start evaluations. Yet, we know little else regarding its effectiveness, particularly over the long-term.1 1

Fisch and Truglio (2001) review the research exploring the impact of Sesa me Street. In terms of short-term effects, Bogatz and Ball (1971) provide a major contribution to this literature, finding substantial improvements in academic achievement within the context of an experimental setting. Diaz-Guerrero, et al. (1976) find similar results regarding the introduction of Sesa me Street in Mexico. Paulson (1974) also uses random assignment, finding that the show improved social outcomes like cooperation. Some research has examined longer-term effects, but that work focuses on differences in outcomes between those who watch versus those who do not, without accounting for the selection into groups (cf. Huston, et al., 2001).

Early Childhood Education by MOOC Kearney and Levine, Page 2

An analysis of the effectiveness of Sesa me Street can potentially also inform current discussions regarding the ability of Massive Open Online Courses (MOOCs) to deliver educational improvements. In essence, Sesame Street was the first MOOC. Although MOOCs differ in what they entail, Sesame Street satisfies the basic feature of electronic transmission of online educational material. Both Sesame Street and MOOCs provide educational interventions at a fraction of the cost of more traditional classroom settings. Most (but not all) MOOCs exist at the level of higher education, which clearly differs from a preschool intervention. Our knowledge of the ability of MOOCs to improve outcomes for its participants is so limited, though, that any proper evaluation of the impact of electronic transmission of educational content is beneficial.2 Additional evidence on the impact of Sesa me Street can also inform a discussion regarding the ability of television to have a positive impact on society. Commentary regarding the role of television largely addresses the potential for negative effects.3 Recent evidence, however, has indicated some beneficial social effects of the introduction of television (Gentzkow and Shapiro, 2008; Jensen and Oster, 2009).4 Other work has shown that rates of teen childbearing fell in response to the MTV series, 16 and Pregnant (Kearney and Levine, 2014).

Sesame Street is another possible example of television that may provide social benefits in the 2

Marcus (2013) raises this issue regarding the lack of evidence examining the effectiveness of MOOCs (Hart, et al, 2015, is a recent exception). Sinha (2011), president of Khan Academy, writes that their own evaluation supports an effect despite their inability to provide experimental evidence. Luzer (2014) reports that an experimental investigation of Khan Academy, funded by the Department of Education, is to be conducted in 2015-16. Whether Khan Academy is a MOOC also remains an open question (Akanegbu, 2013). For our purposes, we are concerned with the general idea of student responses to low cost, electronic educational content and less focused on the specific definition of a MOOC. 3 When Sesa me Street ZDVILUVWLQWURGXFHG%URZQ ZURWH³1RWXQWLOWKHFORVLQJZHHNVRIGLGWHOHYLVLRQ offer a program series that really answered the long-standing criticism of the medium ± namely that it takes a viewer¶VWLPHZLWKRXWJLYLQJDQ\WKLQJLQUHWXUQ± DQGKHOGRXWKRSHIRUDPRUHVXEVWDQWLYHIXWXUH´ 4 Olken (2009), however, presents evidence indicating that access to television in Indonesia was harmful to social capital.


form of improved educational performance, particularly for disadvantaged children. Whether it accomplished this is a matter of evaluating the evidence. The purpose of this paper is to inform these issues by providing further evidence regarding the effectiveness of Sesame Street on early educational performance measures and to extend the analysis to consider longer-term educational and labor market outcomes. Our methodological approach exploits limitations in television technology, which restricted access to

Sesame Street to about two-thirds of the population when it first aired in 1969 (Davis, 2008). As we describe in more detail below, Sesame Street mainly aired on stations affiliated with the Public Broadcasting System (PBS), which often broadcast on UHF (ultra-high frequency) channels. UHF reception was inferior to reception on VHF (very high frequency) channels for physical reasons and because many television sets at that time did not have the capability to receive a UHF signal (McDowell, 2006). Transmission distance also restricted access for some households. Our analysis takes advantage of the county-level variation in YLHZHU¶V DELOLW\ WR watch Sesame Street generated by these technological constraints that existed when the show was introduced in 1969. We combine this geographic variation in broadcast exposure with differences across birth cohorts in terms of their age at the WLPH RI WKH VKRZ¶V LQWURGXFWLRQ Sesame Street¶V FRQWHQW focused on first-grade readiness; those children who had advanced beyond that point would not have been exposed during early childhood and hence would generally not have been affected by its introduction. In the end, the evidence would support an effect of Sesame Street if relative outcomes improved among birth cohorts age 6 and under in 1969 who lived in locations where broadcast reception for the show was high. We implement this approach mainly using data from


the 1980, 1990, and 2000 Censuses.5 We also augment our main analysis, taking advantage of additional data from the 1980 High School and Beyond survey. The results of our analysis provide evidence WKDW6HVDPH6WUHHW¶VLQWURGXFWLRQJHQHUDWHGa positive impact on educational outcomes through the early school years. In particular, exposed cohorts of students with better reception capabilities were more likely to be attending a grade that is appropriate for their age. This effect is particularly pronounced for boys and black, nonHispanic children and those living in economically disadvantaged areas. When we extend the lens farther, however, we are unable to find evidence of substantive improvements in ultimate educational attainment or labor market outcomes. The small estimated impact on wages in adulthood, though, is consistent with forecasts based on the estimated improvements in test scores and grade-for-DJHVWDWXVEURXJKWDERXWE\WKHVKRZ¶VLQWURGXFWLRQ I I. B A C K G R O U N D

A. History of Sesame Street The early 1960s marked a change in thinking among child psychologists and educators, who began to reject the notion that cognitive ability was completely heritable (cf. Hunt, 1961). This sparked interest in early childhood interventions like Perry Preschool and Head Start, which were meant to improve academic preparation among young children. Sesa me Street followed this legacy. It was first proposed in 1967 (Cooney, 1967) and first aired on November 10th, 1969. Its stated SXUSRVHZDV³WRIRVWHULQWHOOHFWXDODQGFXOWXUDOGHYHORSPHQWLQSUHVFKRROHUV´S Following its introduction, Sesame Street was mainly broadcast on PBS channels; of the 192 stations airing the show, 176 of them were affiliated with PBS. The majority of these stations (101) were broadcast on UHF channels rather than VHF channels, which introduced 5

Cascio (2009) uses a similar technique tracking cohorts across successive cohorts in her analysis of the introduction of kindergarten in school districts in the 1960s and 1970s.


technological constraints that limited exposure to the show. As we detail below, only around two-thirds of the population lived in locations where Sesame Street could be received on their televisions. Despite this technological constraint, Sesame Street immediately became a huge success. By January of 1970, over five million households tuned in to a typical episode (Clausen, 1970). Among those between ages 2 and 5, Cook, et al. (1975) estimate that between 28 and 36 percent watched Sesame Street in 1970; between 33 and 42 percent did so in 1971. To put its popularity in perspective, roughly one-third of the entire United States population watches the Super Bowl today. These numbers are even more remarkable within the context of the technological limitations that prevented around one-third of the population from watching it.

Sesame Street ZDVUHDOO\WKHILUVWFKLOGUHQ¶VWHOHYLVLRQSURJUDPPLQJH[SOLFLWO\IRFXVHGRQ educational content. Popular shows among children at that time included Captain Kangaroo (CBS), Mister Rogers Neighborhood (PBS), Romper Room (locally produced), cartoons (e.g., The Jetsons, The F lintstones), and sitcoms (The Andy Griffith Show). Earlier shows prior to

Sesame Street that explicitly targeted children include Looney Tunes, Tom and Jerry, Howdy

Doody, and Kukla, F ran, and Ollie, among others. Some shows, like Mister Rogers Neighborhood, did focus on teaching social skills, such as getting along with others, but none focused on academic content in the way that Sesa me Street did. It is also important to recognize that at the time Sesame Street was introduced, preschool attendance was the exception, rather than the norm. According to our calculations using the 1970 Census, only 9 percent and 19 percent of children ages 3 and 4, respectively, attended


preschool.6 Among children age 5, 57 percent attended school (presumably kindergarten), but, at the time, kindergarten generally only lasted half a day. Many of those children would still have had the chance to watch Sesame Street (if they lived in a home that could receive the necessary broadcast signal) after school, since stations typically broadcast the show both in the morning and later in the afternoon.7

B. Television Technology/History of U H F Limited access to Sesa me Street often generated by UHF transmission is a critical component of our analysis.8 For the television viewer, the most straightforward distinction is that VHF channels are those between 2 and 13 and UHF channels are those greater than 13. Signals broadcast over VHF travel farther and their reception is less affected by mountains, buildings, and other obstacles. A household with a television capable of receiving both signals is more likely to be able to receive a VHF signal than a UHF signal. The second limitation to access is based on the ability of television sets at the time to receive UHF signals. Before 1952, there were no UHF channels broadcasting, initially because there were enough VHF channels to meet demand and, subsequently, because of World War II (the military wanted the rights to those frequencies) and a resulting Federal Communications Commission (FCC) moratorium on granting new station licenses.9 When the moratorium was 6

These estimates are comparable to those provided by Gibbs, Ludwig, and Miller (2013). Cascio (2009) describes the institutional background regarding the introduction of kindergarten across school districts in the United States, detailing its limited availability during this period. 7 Based on 1980 Census data that we describe below, we calculate that 29 percent of children lived in locations where Sesa me Street was broadcast only in the morning, 11 percent only in the afternoon, and 59 percent both in the morning and in the afternoon. We experimented with taking advantage of this variation in the data, but found insufficient power to identify differential effects. 8 Webbink (1969) and Rothenberger (2004) provide useful discussions of the history of UHF broadcast technology and related public policy. 9 After World War II, in response to strong demand for new channels and the need to insure that signals from neighboring communities would not interfere, the moratorium was imposed to develop a viable licensing system. 7KLVPRUDWRULXPLVDQLPSRUWDQWHOHPHQWLQ*HQW]NRZDQG6KDSLUR¶V DQDO\VLV of preschool age exposure to WHOHYLVLRQRQFKLOGUHQ¶VVWDQGDUGL]HGWHVWVFRUHV.


lifted in 1952, strong demand for additional stations led the FCC to expand channel options to include those broadcasting in UHF. Even then, television manufacturers typically did not build sets that included UHF tuners. With no tuners, original programming was limited; with no original programming, there was no demand for the tuners. This changed when the ³All Channel Receiver Act´ became law in 1962. This act, which took effect in 1964, required manufacturers to produce television sets that could receive both UHF and VHF signals. Televisions were expensive, though, costing roughly $700 for a state-ofthe-art 25-inch color set in 1966 (TV Guide, 1966), compared to the median family income of $7,436 in that same year (U.S. Bureau of the Census, 1968). This resulted in a slow diffusion of UHF signal receipt. In 1969, 95 percent of households owned television sets, but only 54 percent of those households had one that could receive a UHF signal (U.S. Bureau of the Census, 1970). We describe below how we use these constraints to identify the effect of potential exposure to

Sesame Street SURJUDPPLQJRQFKLOGUHQ¶Vsubsequent outcomes.

C. ETS Study of Sesame Street¶V,PSDFW Educational Testing Service (ETS) was hired right from the initial development of

Sesame Street to design and implement an analysis that would examine whether children who watched the show performed better on a range of educational activities (Bogatz and Ball, 1971). Ex-ante, WKH GHVLJQ XVHG LQ WKH ILUVW \HDU IROORZLQJ WKH VKRZ¶V LQWURGXFWLRQ was a good one. Disadvantaged children in a number of locations were identified and randomly assigned to treatment and control groups, where treatment included explicit encouragement to watch the show.10 Clearly, the evaluators had no idea Sesa me Street viewership would be so high. The 10

The component of encouragement is one that Cook, et al. (1975) criticize, arguing that doing so can highlight the importance of academic skills, which can have an impact on its own even in the absence of Sesa me Street content.


experimental design failed because so many members of both the treatment and control groups watched it. ETS adopted a different approach in the second year evaluation. This evaluation is based on experiences in Winston-Salem, NC and Los Angeles, CA and relies on viewing constraints among low-income households. In Winston-Salem, NC, cable television (in its infancy at this time) was just being introduced, enabling subscribers to watch Sesa me Street in a location where reception would have been limited otherwise. But low-income households were unlikely to be able to afford cable service. ETS randomly assigned families with pre-school age children to control and treatment groups, where treatment group families were provided with cable television and the children in those households were encouraged to watch the show. In Los Angeles, Sesame Street was available on a UHF channel only and many low-income households did not have a television that could receive a UHF signal. In that location, ETS provided treatment group households with UHF converters for their television, enabling them to watch the show; again, children in treatment group households were encouraged to watch it. The results of this analysis provided support for a strong effect of Sesame Street. Positive effects were observed on a number of specific outcomes. Among their many results, we focus on the impact of the show on the Peabody Picture Vocabulary (PPVT) test, which is standardized and frequently used as a general test of cognitive performance. The results of this exercise indicate that the treatment group experienced a 0.36 standard deviation relative increase in PPVT scores.11 This effect can be interpreted as the equivalent of around an additional full year of

11

%RJDW]DQG%DOO UHSRUWWKDWWKH³HQFRXUDJHG´JURXSH[SHULHQFHGJDLQVLQ3397VFDOHGVFRUHVRISRLQWV EHWZHHQWKHEHJLQQLQJDQGHQGRIWKH \HDU0HPEHUVRIWKH ³QRWHQFRXUDJHG´ JURXSVDZWKHLUVFRUHV IDOOE\ -3.7 points (see p. 105). Cook, et al. (1975) reports that the PPVT is designed to have a mean of 100 and a standard deviation of 15. The difference between the two groups of 5.4 points reflects 0.36 of a standard deviation increase.


learning.12 This is an important study and the evidence is compelling. But Cook, et al. (1975) note in their critique of the ETS study, ³we could find no data of any reasonable quality to assess the crucial question of the long-WHUP HIIHFWV RI YLHZLQJ µSesame Street¶ RQ ERWK Oearning and VRFLDOGHYHORSPHQW´ This has remained a gap in the literature, which we attempt to fill in with this paper.

D. Sesame Street and Head Start Along many dimensions, Sesame Street is a comparable intervention to Head Start. Both shared a common heritage, incorporating new ideas during the 1960s about the role that early childhood interventions can play in improving educational outcomes for disadvantaged children. Both had demonstrated evidence of short-term effectiveness. Gibbs, et al. (2013), Duncan and Magnuson (2013), and Shager, et al. (2014) report that the literature assessing the impact of Head Start on early childhood cognitive test scores yield similar effect sizes to those reported in Bogatz and Ball (1971) regarding Sesame Street. In fact, policy makers and analysts who were focused on early childhood education at the time specifically compared the two programs, noting that Sesame Street appeared to generate improvements in cognitive skills comparable to those of Head Start for a fraction of the cost.13 It is important to keep in mind, though, that the two interventions are actually substantively different from each other, not only in the nature of the intervention, but in their stated goals. Head Start was designed to be a comprehensive program with a variety of services 12

This translation is based on Hill, et al. (2008), which indicates that an effect size of 0.20 on nationally normed standardized tests is roughly equivalent to six additional months of learning. 13 =LJOHUDQG0XHQFKRZ VWDWHWKDW1L[RQDGPLQLVWUDWLRQRIILFLDOVDUJXHG³we can get Sesa me Street to reach poor kids by spending sixty-ILYHFHQWVSHUFKLOG«:K\VKRXOGZHVSHQGRYHUDWKRXVDQGGROODUVSHUFKLOGRQ+HDG Start?´ (p. 165). Cook, et al. (1975) similarly state: ³why pay teachers more money, run programs such as Head Start, build new schools, and experiment with complex teaching machines when there already is a relatively LQH[SHQVLYHHDVLO\H[SDQGDEOHDSSURDFK"«&OHDUO\XQOHVVWKHµSesa me Street ¶DSSURDFKUHDOO\LVHIILFDFLRXVWKHUH is the concern that the swell of enthusiasm for it could drown out other educational efforts to improve human UHVRXUFHV´p. x in the foreword).


beyond typical preschool education, including, for example, medical and dental services and family outreach. It certainly included D FRPSRQHQW GHVLJQHG WR LQFUHDVH FKLOGUHQ¶V FRJQLWLYH development as Sesa me Street did, but it went far beyond that to address social and emotional development, improve self-confidence, address health deficiencies, and improve family functioning (Cooke, 1965). The initial goals of the Sesame Street program were more narrowly focused, extending beyond academic achievement only to address cultural development (music and the arts) and the awareness of basic emotions, including aggression and fear (Cooney, 1967).14 Because of these fundamental design differences it is not clear that we would expect similar long-term effects from the two programs. I I I. M E T H O DS A N D D A T A

A. Overview As described earlier, only two-thirds of households reportedly were able to receive the signal broadcasting Sesa me Street when the show began in 1969. The general framework of our empirical approach is to determine the variation in ³FRYHUDJH´ OLNHOLKRRG RI EHLQJ DEOH WR receive the signal) across locations (counties). We then examine whether outcomes improve as coverage improves for those children who were of preschool age in 1969 relative to those who had already started school at that time. The latter group should have received little or no effect from the show. Coverage rates, though, are potentially related to the characteristics of households in a county. Wealthier households, for instance, would be better able to afford a new television that could receive UHF signals following the 1964 law change. To overcome this 14

Edward Zigler, Director of the Office of Child Development (which oversaw Head Start) in the Nixon $GPLQLVWUDWLRQUHVSRQGHGWRWKH:KLWH+RXVH¶VFRPSDULVRQof Sesa me Street DQG+HDG6WDUWE\VWDWLQJ³Kow long would a poor child have to watch Sesa me Street to get his or her teeth filled? When nobody could answer, that was WKHHQGRIWKHPHHWLQJ´ (Zigler and Muenchow, 1994). The point was that Head Start did much more than just try to impact educational skills.


problem, we implement an instrumental variables strategy where our instruments for coverage rates represent the distance to the closest television tower broadcasting Sesame Street and whether that tower transmits using UHF or VHF. Since the location of the television towers and the channel they were assigned are functions of FCC decisions made years before Sesame Street began, they should be unrelated to any subsequent changes in outcomes that occurred after

Sesame Street was introduced. The remainder of this section provides the details regarding our implementation of this approach.

B. Calculating Sesame Street Coverage Rates A critical data component of our empirical strategy is a measure of Sesame Street broadcast coverage rates across geographical areas. We need to know: (a) which television stations broadcast Sesame Street in a given area and (b) what share of households in a given area FRXOG UHFHLYH WKDW VWDWLRQ¶V EURDGFDVW VLJQDO 'DWD IURP &KLOGUHQ¶s Television Workshop (undated) indicate which stations aired Sesame Street in 1969/1970. Estimating what share of households in a given area could receive the signal from that station requires that we impute coverage based on supplementary data sources. For this purpose we rely heavily on data reported in the 1968-1969 edition of a trade publication, TV F actbook (Television Digest, Inc., 1968-1969).15 This data source provides a listing of every television station ± including both commercial and non-commercial ± broadcasting in the United States, along with its technical specifications: channel number (which captures UHF/VHF), latitude and longitude of its broadcast tower, height of the tower, and transmission signal power. For all commercial stations, the publication also lists surrounding counties and coverage rates, as defined by the fraction of television households who have the 15

Data Appendix I gives more details about our data procedure to estimate coverage rates.


ability to receive the signal from each station.16 Using this sample, we regress county-level coverage rates on the technical specifications of each channel ± including UHF/VHF status, distance between the broadcast tower and the population centroid of each county, transmission power, and height of the tower ± to establish the empirical relationships between those specifications and the coverage rate among commercial stations. We then apply those estimated relationships to the specifications of the non-commercial stations to obtain an estimated coverage rate at the county level for those stations. We assign to each county the station that provides the highest coverage rate and use that rate as our measure of the percentage of households in that county that can watch Sesame Street in their homes. In the end, our process generates a national coverage rate of 65 percent, which aligns closely to national estimates of Sesame Street coverage made at that time.17 The regression results from estimating coverage as a function of station specification make clear the importance of UHF versus VHF in determining coverage. Coverage rates for a UHF station in the same county as the television tower are 42 percentage points lower than they are for a comparable VHF station in that county. This roughly corresponds to the limited number of televisions that could receive UHF signals at that time. Coverage in the county falls around another 20 percentage points for a UHF station if it is 100 miles away from the tower. Figures 1 and 2 display all the stations that broadcast Sesame Street, distinguishing them by VHF/UHF status, and county-level coverage rates. Figure 1 indicates that Southern 16

Reported coverage rates are categorical (greater than 50 percent, 25 to 50 percent, and 5 to 24 percent). We linearize these rates, assuming the categories reflect 90 percent, 40 percent, and 20 percent coverage, respectively. We have also experimented with alternative assumptions and non-parametric specifications, but the results based on these alternatives were consistent with our linearized approach. We focus on this approach because of the greater simplicity in estimation and interpretation. 17 Clausen (1970) reports that the coverage rate was 69.4 percent the week of January 19-25 of 1970. The coverage rate data that we use refers to 1968-69, which is likely slightly lower than this level because some households upgraded their television sets over this interval, acquiring the ability to view UHF signals.


California, Ohio, and the District of Columbia represent examples of heavily populated areas where Sesa me Street was only accessible via a UHF signal. Figure 2 shows that these areas have very low rates of coverage as well, indicating that UHF transmission substantially reduced access to the show. We also compare our coverage rates to additional data on Sesame Street Nielsen ratings to provide further evidence that our measure of coverage adequately captures constraints on audience size. Local ratings data for children between the ages of 2 and 5 when Sesame Street was introduced are available for 28 metropolitan areas (Haydon, 1973).18 We link these areas to our constructed county level coverage data. Figure 3 presents a scatter plot of this relationship. It shows that coverage and ratings are positively related; the correlation coefficient between them is 0.48. The slope also provides informative data. For every percentage point increase in coverage, ratings among 2 to 5 year olds increase by 0.58 percent. If preschool age children are proportionally distributed across counties, this result would suggest that over half of the children who have the technological capability to watch Sesame Street watch the show. One potential problem with estimating the relationship between subsequent outcomes and coverage rates is that differences in coverage may not solely be a function of technology. They may also be affected by the characteristics of households in a county.19 Consider two counties the same distance from a television tower broadcasting Sesame Street over UHF, one of which is considerably wealthier than the other. The higher income county likely has more households with televisions that can receive the UHF signal, increasing its coverage. The lower income 18

The data in Haydon (1973) report the number of viewers between the ages of 2 and 5 who watch Sesa me Street in each reported metropolitan area. We combine these data with population counts from the 1970 Census for this age group to generate the percentage of viewers. 19 The 1970 Census includes a question regarding whether the household owns a television set that can receive UHF signals. We have estimated regression models of UHF signal receipt on household characteristics. Family income, greater education, and younger age are all strongly positively correlated with living in a household owning a television that can receive UHF signals.


county may be more restricted in its UHF coverage and have households more likely to rely on signal transmission from a more distant VHF channel, decreasing its coverage. To overcome this problem, we implement an instrumental variables strategy. Our instruments consist of the distance to the closest television tower broadcasting Sesame Street (not necessarily the one with the highest coverage) and an indicator for whether that tower transmits over UHF. This form of identification strictly relies on television technology, along with historical FCC decisions regarding placement and transmission frequencies from television towers, not household characteristics.

C. E mpirical Specifications Our main identification strategy relies on variation across cohorts and geography (counties) in FKLOGUHQ¶Vpotential exposure to Sesame Street programming. We focus on cohorts born between 1959 and 1968, who would range in age from 1 to 10 years of age when Sesame

Street was introduced in 1969. Another way to think about this cohort variation in exposure to the show is that individuals born between 1959 and 1963 would have been age 6 or older, and already in elementary school, at the time the show first aired. Individuals born between 1964 and 1968 would have been age 5 and below, and would have been exposed to the program during their preschool age years. The other dimension of variation in access to the show comes from the county-level Sesame Street coverage rate, defined (as described in detail above) to be the share of television households in the county who were able to receive a signal over which Sesame

Street was broadcast. Our empirical strategy is to observe whether outcomes differ for cohorts of individuals who would have had preschool-age exposure to the show as compared to older cohorts and individuals who lived in counties with limited broadcast coverage. Any difference along these


dimensions could reasonably be considered to be evidence of an impact of the show on educational outcomes. Within this framework, the specific models we estimate take the following forms: O u tco m e i jc

E 0 E 1 * p r eschoo l 69 ic * SS C ov j E 2 P o l i cy jc E 3 X i jc J c * J s G j H i jc 1973 / 74

O u tco m e i jc

E0

¦

E c * J c * SS C ov j E 2 P o l icy jc E 3 X i jc J c * J s G j H i jc

(1) (2)

c 1967 / 68

where the dependent variable, Outcomeijc, represents different educational and labor market outcomes for individual i, in county j, in cohort c. The subscript s indexes states. Sesame Street coverage rates, SS Covj, are only indexed by county; that variable is time-invariant and captures the technological constraints to watching the show that existed just before it was introduced. In Equation (1), we estimate a difference-in-difference specification where E1 captures the causal effect of interest. Birth cohorts are distinguished by those who would be preschool age when the show began in the fall of 1969 (preschool69ic). We include county fixed effects (ߜ௝ ) to capture time-invariant differences in outcomes across counties. We also include state*birth cohort fixed effects (Jc*Js) to capture time varying changes in outcomes across states, including state-level policy variation like the introduction of Medicaid and welfare generosity. The model also controls for a series of individual covariates, captured by X ijc: race/ethnicity, age, and, when it is available, socioeconomic status. We instrument preschool69*SS Cov with distance to the closest television tower broadcasting Sesame Street and whether that tower broadcasts UHF signals, both interacted with preschool69. A potential source of omitted variable bias would occur if other policies or environmental conditions that would affect the educational outcomes of young children changed at the county level around the same time as Sesame Street was introduced. Two potential candidate policy


changes that could lead to an omitted variable bias include the introduction of the Food Stamp program and increased expenditures on the Head Start program. The Food Stamp Program was introduced across counties in the U.S. between 1961 and 1975. Our model includes an indicator for having the Food Stamp Program in operation in the county in a given year, as measured by Hoynes, Schanzenbach, and Almond (2012). Our model also controls for Head Start expenditures in 1968 and 1972, as recorded by Ludwig and Miller (2007).20 In Equation (2), we relax the pre-assignment of treatment and control groups by cohort, allowing for the effect of Sesame Street coverage to differ by grouped birth cohorts. We aggregate individuals from neighboring birth years to increase the power of this analysis. This SURYLGHVDVSHFLILFDWLRQ FKHFNWR VHHLIWKH ³WUHDWPHQWHIIHFW´UHDOO\VWDUWVZLWK WKHDSSURSULDWH groups of birth cohorts. Again, we instrument the cohort*coverage interactions with cohort*distance and cohort*UHF status of the closest television tower. In both cases, we also estimate models separately by race/ethnicity and gender.21 Our primary measure of interest is access to Sesame Street broadcasting, not actual viewership of the show. In this sense, our approach identifies an ³intent to treat´ relationship, not a ³treatment on the treated.´ If we had better ratings data, we could pursue an approach that would help us address the impact on the marginal viewer. As noted earlier, though, ratings data are only available for a limited subset of counties that are encompassed by 28 metropolitan 20

We are grateful to Diane Schanzenbach for providing us with this Food Stamp program indicator. We are also grateful to Jens Ludwig for providing the Head Start data. Those data provide Head Start expenditures by county in 1968 and 1972; we linearly interpolate to fill in values between the two years. Since Head Start largely enrolls fouryear-olds and school entry is at age 6, we treat these data as affecting 6 year olds in 1970 through 1974. We linearly interpolate to fill in values between the two years. 21 We are unable to disaggregate individuals by socioeconomic status because of data limitations. The only feasible measure of SES is maternal education, but that only exists in our 1980 Census subsample for those individuals still living with their mothers. We include maternal education along with a missing value indicator for those who no longer live with their mother as explanatory variables, but the sample selection issues introduced by splitting the sample in this way are more difficult to overcome.


areas.22 As a result, the question that we are best able to address with our analysis is what is the impact of making a show like Sesame Street more readily available to children, not what impact does it have on an individual child watching the show.

D. Overview of Census Data Our main sources of individual-level data used to estimate equations (1) and (2) are microdata from the 1980, 1990, and 2000 U.S. Census of the Population, available from IPUMSUSA (Ruggles, et al., 2010). We restrict our Census samples to individuals born between 1959 and 1968; these individuals would have entered first grade from around 1965 until 1974 (depending on exact birthdates and state laws regarding age of school entry), just before and after

Sesame Street was introduced. Table 1 eases the interpretation of cohorts advancing through these censuses, indicating points in the lifecycle of those in the analysis sample as they age. First, we use the 1980 Census to examine elementary school performance as captured by ³grade-for-age´VWDWXV; this measure indicates whether a child is enrolled in school ± or graduated high school for those over age 18 ± at a grade appropriate for his or her age.23 We use data on quarter of birth available in the 1980 Census along with the school entry requirement laws listed in Cascio and Lewis (2006) to refine our estimates of the year students would be expected to start school.24 The birth cohorts in our 22

We have experimented with using an instrumental variables strategy to estimate this model with these restricted data ± instrumenting for ratings with coverage ± but we found that the small sample size reduced the statistical power of the analysis to a point where it is not informative. 23 Grade-for-age status reflects a stock; falling behind could have taken place several years ago. Our analysis of data from the 1980 Census of those born between 1962 and 1974 (school-age at that time) indicates that many students fall behind in first through fourth grades. As Deming and Dynarski (2008) document, purposely holding children EDFNIURPVWDUWLQJ ILUVW JUDGH DWDJH VL[³UHG-VKLUWLQJ´ LV D PRUH recent phenomenon. Our data on grade-for-age status is consistent with that reported by Hauser (2004). Cascio (2005) indicates that grade-for-age is an imperfect proxy for grade repetition. 24 Where necessary, we make the most generous assumptions possible to increase the likelihood of a child being in the appropriate grade level. For instance, a child born in the 3 rd quarter in a state where children are supposed to enter 1st grade in September of the year they turn age 6 is treated as if he or she is still age 5 in that year; his or her birthday could have been in August.


sample would be between the ages of 12 and 21 in the 1980 Census. Second, we use the 1990 Census to measure ultimate educational attainment (high school dropout, high school graduation, or attended any college), as these birth cohorts would be between ages 22 and 31 in 1990 (it is not possible in the 1990 Census to distinguish high school graduation from GED attainment). And third, we use the 2000 Census to measure labor market outcomes (employment, hourly wage, and poverty status) when these birth cohorts would have been between the ages of 32 and 41, presumably established in the labor market, if working. An important issue to address regarding the use of Census data is migration. Ideally, we would know the state and county of residence in which an individual resided in 1969, when

Sesame Street began, but this information is not available. For these birth cohorts, a reasonable alternative would exist if we knew the state and county of birth since they would all be very young in 1969 and mainly living in the same place. State of birth is, in fact, available in these Censuses, but county of birth is not. To circumvent this problem we restrict our samples to those individuals whose recorded state of residence in the Census is the same as their state of birth. This sample restriction assumes that interstate mobility since birth is unrelated to Sesame Street coverage in 1969. For this subsample, we assign county of residence in the Census year to be the county of birth. The assumption that we maintain in this assignment is that individuals who live in the same state where they were born are likely to remain in a county close to where they were born and, importantly, in the same television market with similar broadcast reception. Table 2 provides some evidence suggesting that our sample decisions regarding migration are reasonable. For this purpose, we first examine mobility data from the Census and compare it to the 1979 National Longitudinal Survey of Youth (NLSY79), which provides more


geographic detail. Respondents in the NLSY79 were born between 1957 and 1964, comparable to the birth cohorts we examine with Census data (1959 to 1968). First, we explore interstate mobility between birth and 1980, 1990, and 2000 using the Census, and then using NLSY79 interstate mobility between birth and the same years. The results in Table 2 provide evidence indicating interstate mobility in the Census and NLSY79 are quite comparable; around 77 percent, 67 percent, and 65 percent of respondents lived in the same state in 1980, 1990, and 2000, respectively, in the two datasets. If we restrict the NLSY79 sample in each year to those who lived in the same state as they were born, we also see that over 80 percent lived in a county within 60 miles of their birth county. We also used the NLSY79 data to estimate regression models comparable in format to that described by Equation (1) except that the dependent variable is an indicator of state outmigration. The results reported in the bottom panel of Table 2 do not provide evidence of selective state outmigration between birth and 1980 or 1990. Those young enough to have seen

Sesame Street when it first came out in areas where more residents could see it are not significantly differentially likely to live in a state different from where they were born. We also find no evidence of a relationship between Sesa me Street exposure and movement to a county outside a 60 mile radius from the county of birth for those who never left their home state. Those who did move, however, will generate some attenuation bias in our Census analysis resulting from measurement error in Sesame Street exposure. We will address this issue further when we describe the results of our analysis. Using Census data to estimate our equations of interest requires one additional sample restriction. County of residence is only identified for those individuals in sufficiently heavily populated counties. This eliminates roughly half the sample. Those included in the 1980 Census


live in just 349 counties, for instance. The location of these counties is documented in Figure 4, which demonstrates the focus on more heavily populated areas. A beneficial result of this data constraint is that identification in our statistical model is not driven by comparisons of urban and rural locations. In fact, using data from the 1970 Census, we show in Table 3 that among counties with available data in the 1980 Census, those distinguished by the quality of reception are generally quite similar.25 Those counties not in the sample are quite a bit smaller and poorer than the others. This makes sense since the reason they are not separately identified is because they are too small, and, at the time, rural poverty was an even greater issue. Differences, though, between strong and weak reception counties (defined below) among those separately identified are rather limited; t-tests comparing them are unable to distinguish the observed differences from random variation. Certainly this analysis does not prove that these groups are randomly selected, but it does indicate that no obvious selection differentiates these two sets of counties. I V . R ESU L TS

A. Graphical Analysis Before presenting our formal results from estimating Equations (1) and (2), we first present Figures 5 through 8, which are designed to illustrate our identification strategy and guide the interpretation our subsequent findings. For this analysis, we distinguish counties by their distance to the closest tower broadcasting Sesame Street and whether that broadcast was UHF or 9+):HGHILQH³VWURQJUHFHSWLRQ´FRXQWLHVWREHWKRVHZLWKLQPLOHVRIRQHRIWKRVHWRZHUV DQGEURDGFDVWLQJLQ9+)$OORWKHUFRXQWLHVDUHGHILQHGWREH³ZHDNUHFHSWLRQ´FRXQWLHVEHFDXVH they violated at least one of those conditions. Based on our estimated coverage rates and data 25

These data come from the 1972 County and City Data Book, which is available from ICPSR at: (http://www.icpsr.umich.edu/icpsrweb/NACDA/studies/61, accessed 4/23/15.


from the 1980 Census, these categories roughly split our sample in half, with coverage rates of 83 and 55 percent, on average, in strong and weak reception counties. We then trace out differences in our outcome measures (grade-for-age status, educational attainment, and labor market outcomes) by exact cohort between those who live in strong and weak reception areas. This approach mimics a simplified reduced form version of the analysis we report below. Here we aggregate counties¶ reception capabilities rather than use them continuously in our full econometric analysis. Figure 5 compares grade-for-age status across these categories for several birth year cohorts. The horizontal axis is distinguished by cohort and is designed to identify those who may have been affected by Sesame Street¶V introduction. In the 1980 Census, quarter of birth data is available and we use that to better determine school start year. In other Census years, we rely on the year a child turned age 6. If Sesame Street had an effect on educational or labor market outcomes, we would expect to the see the effect for the birth cohorts who started school (in the 1980 Census) or turned 6 (in the 1990 and 2000 Census) in 1970 or later. These children would have been preschool age when the show was introduced. Those who started school or were older than age 6 in 1968 or earlier would not have been as directly affected by its introduction, since the show targets preschool age children and academic skills most relevant to school entry. We view those individuals who were supposed to start school in 1969 ± those who were typically age 6 at the time ± as a transitional cohort. Although a show designed to improve first grade readiness should not have a large impact among those who were already in first grade, some children in this cohort may not have started school yet. Others may have started school, but were not doing well in first grade. Since Sesame Street was shown in most locations during after school hours, these students may have benefited from its availability as well. We do not expect


any effect on them to be as large as for the younger cohorts, but we recognize the possibility of an intermediate effect. The results of this simple analysis regarding grade-for-age status reported in Figure 5 support the hypothesized patterns. For cohorts who should have started school in 1968 and earlier, there is not a large difference in grade-for-age status between those in stronger versus weaker reception counties. For cohorts who should have started school in 1970 and later, there is a clear positive difference in rates of grade-for-age between those in the two areas. As a whole, those in the strong reception counties are 1.5 to 2 percentage points more likely to be at the grade level appropriate for their age in 1980. Differences in grade-for-age status for the 1969 school start year cohort are positive, but smaller than those for subsequent cohorts, as expected. Overall, this figure provides evidence supportive of an effect on grade-for-age brought about by exposure to Sesame Street. :HH[WHQGWKLVDQDO\VLV WR FRQVLGHUZKHWKHUWKH VKRZ¶VLPSDFW LV JUHDWHVW DPRQJPRUH disadvantaged children. As we described earlier, however, we do not have a good way to measure socioeconomic status in the census data available to us. Instead, we augment the census data with additional data from the 1970 Census distinguishing locations by different levels of disadvantage. This approach focuses on the relative disadvantage of an area, not the individual. In this analysis, we focus on the percentage of the county population that has less than a high school degree, splitting counties by those that are above and below the median value.26 The results of this analysis, reported in Figure 6, indicate that the impact of Sesame Street¶V introduction on grade-for-age status is considerably greater in more disadvantaged areas. Cohorts 26

In subsequent analyses, we also use these data to distinguish counties by the percentage of families headed by a female head, the percentage of families with incomes below $5,000 in 1970, and the percentage of the population that is black. Results for the present analysis using high and low values of these other measures are comparable to those reported here.


who were of pre-school age at the time the show began were roughly 3 percentage points more likely to be at the appropriate grade level for their age if they lived in an area with strong reception. This is perhaps twice the estimated effect for the population as a whole. The results regarding educational attainment and labor market outcomes, reported in Figures 7 and 8, are less compelling. We chose the likelihood of graduating high school (including those who attend college) as our indicator of educational attainment and the SHUFHQWDJHRIWKHSRSXODWLRQWKDWLVHPSOR\HGWKH³HPSOR\PHQWUDWH´ as our indicators of labor market performance. In neither case is there anything resembling a positive break from trend around the 1969 school entry cohort, defined by the year a child turned age 6.

B. Econometric Results: Grade-for-Age Table 4 presents our estimation results reflecting the models represented by Equations 1 and 2 above, implemented using the IV strategy described earlier.27 For the full sample, from the top panel of the table, reporting results from estimating Equation 1, we find that children who were preschool age in 1969 and who lived in areas with greater predicted Sesame Street coverage were statistically significantly more likely to be at the grade level appropriate for their age. To interpret the magnitude of the coefficient, we consider the impact of a 30 percentage point increase in coverage rates, which is tantamount to moving from a typical area with weak reception to an area with strong reception. The hypothetical case of moving from no coverage to complete coverage is out of sample and thus an inappropriate basis of comparison. A 30-point increase in coverage rates would generate a 2.9 percentage point (0.3*0.097 = 0.029) increase in 27

Appendix Table 1 presents estimates on the excluded instruments from the first stage regressions in models that aggregate the treatment effect across all cohorts of preschool age starting in 1969 by demographic subgroup using the 1980 Census. As exemplified by these results, all first stage estimates are very strong. We only report these to limit the number of results that all show essentially the same thing. Appendix Tables 2 through 5 report the results of OLS estimates of the same models.


the rate of grade-for-age. With 20.3 percent of the sample behind their appropriate grade in school, this estimate implies that moving from a weak to strong reception county would lower that rate by around 14 percent.28 This effect on grade-for-age status is particularly pronounced among boys and black, non-Hispanic children. For boys, grade-for-age status is estimated to jump 3.8 percentage points (0.125*.3) between weak and strong reception counties. This reflects a similar 16 percent reduction in the likelihood of being below grade level, because this is one of the groups that has the most room for improvement in grade-for-age status. For black, non-Hispanics, grade-for-age status is similarly estimated to jump by 4.1 percentage points (0.136*.3), which also reflects a 13.7 percent reduction in being behind in school. The impact for white, non-Hispanics is still sizable; grade-for-age status is estimated to rise by 1.4 percentage points and close the gap by 8.3 percent.29 Coefficient estimates by gender and race are significantly different from each other. As a point of comparison, Currie and Thomas (1995) report estimates from sibling difference models indicating that white children are 47 percent less likely to repeat a grade if they attended Head Start relative to a sibling who did not. They did not find a statistically significant effect for African-Americans. We find that white and black children who were age 5 or younger when Sesame Street was introduced and who lived in an area with strong reception were 8.3 and 13.7 percent, respectively, less likely to be below the grade level appropriate for their age relative to children in weak reception areas. Recall that coverage rates are 28 percent higher in strong reception areas than in weak reception areas (83 versus 55 percent). For purposes of comparison we simulate what would happen if we moved from no coverage to 28

The fact that this estimate is somewhat larger than what one would expect from Figure 5 is attributable to including state*birth cohort fixed effects. In a more traditional difference-in-difference analysis including birth cohort and county fixed effects, our results are very similar to what we observe in Figure 5. 29 Estimates for Hispanics are difficult to interpret because of the weaker power available with a smaller sample size.


complete coverage (admittedly an out-of-sample prediction). This generates the prediction that the rate of being behind grade-level in school falls 29.6 percent for white, non-Hispanics and 49 percent for black, non-Hispanics. These results suggest that the introduction of Sesame Street for white and black, non-Hispanic children had similar effects on elementary school performance as did participation in the Head Start program. The bottom panel of the table reports the results of the more descriptive model in Equation 2. To obtain more power, we aggregate birth cohorts into two year intervals. Because we want the interaction between coverage and the 1969 school start year cohort to enter the model separately as a possible transition year, we move 1970 to form a 1970 through 1972 aggregated birth cohort. Again, the results of this analysis strongly correspond to the patterns reported in Figure 5. For those scheduled to start school prior to 1969, we see no significant difference in outcomes. For those cohorts that started school in 1970 and afterwards, we see a statistically significant (at least at the 10% level) increase in grade-for-age status associated with greater Sesa me Street coverage. For the 1969 school start year cohort, we see an effect between the younger and older groups, as hypothesized. These results support the interpretation of a causal effect. Again, the estimated impact is particularly large for boys and black, non-Hispanic children. :HH[WHQGWKLVDQDO\VLV WR FRQVLGHUZKHWKHUWKH VKRZ¶VLPSDFW LV JUHDWHVW DPRQJPRUH disadvantaged children. As we described earlier, however, we do not have a good way to measure socioeconomic status in the census data. Instead, we augment the census data with additional data from the 1970 Census obtained from the 1972 County and City Databook, which distinguishes locations by different levels of disadvantage. This approach focuses on the disadvantage of the area, not the individual. The indicators we use are the percentage of the


county population that has less than a high school degree, the percentage of families headed by a female head, the percentage of families with incomes below $5,000 in 1970, and the percentage of the population that is black. The top row of Table 5 indicates mean values for these indicators. The remainder of Table 5 presents regressions results that are largely of the same form as Equation 1, except that they also introduce an additional interaction term between those who would be of preschool age after 1969 in counties with different predicted Sesame Street coverage rates and the four measures of economic disadvantage in the counties. 30 In each case, we find that the ability of Sesame Street to increase the rate at which children are at grade level is greater in areas characterized by greater disadvantage. For instance, a county where two-thirds of the households are headed by high school dropouts (compared to the mean rate of 44.6 percent) would experience about a 6.3 percentage point increase in grade-for-age status if Sesame Street coverage rose from a low level to a high level [0.3*(66.7*.0056-.165)]. This compares to the average effect of 2.9 percentage points calculated earlier.

C. Econometric Results: Educational Attainment and Labor Market Outcomes We now move on to consider later life outcomes, including ultimate educational attainment and labor market outcomes. We estimate completely analogous models to those reported earlier other than changing the dependent variable and the specific Census data on which we estimate the models. The measures of educational attainment (high school dropout, high school graduate, or attending any college) come from the 1990 Census. Measures of labor

30

The reader should note that this is a restricted version of a more general triple-difference estimation strategy, as described in Gruber (1994). The main difference is that our model restricts the coefficients on the interactions between cohort fixed effects and county characteristics to be zero. When we relax this restriction, we find that our data are not sufficiently powerful to separately identify all of the effects without a sizable impact on the standard errors of our variables of interest. With the restriction imposed, bias would result if children from more disadvantaged counties experienced differential educational outcomes well-timed to the introduction of Sesame Street. This is an unlikely, but not impossible scenario.


market outcomes (log hourly wage, employment status, and poverty status) come from the 2000 Census. The results of this analysis are reported in Table 6; the top panel reports models of the form represented in Equation (1) and the lower panel breaks out these effects by specific birth cohort, as described by Equation (2). Estimates are reported for each educational attainment and labor market outcome for the full sample (each cell represents the results from a separate regression). The results in the first three columns regarding educational attainment provide no evidence of changes in these outcomes. Parameter estimates are small, statistically insignificant, and inconsistent with the expected pattern across cohorts (in the bottom panel).31 The results in the last three columns regarding labor market outcomes are less clear. Parameter estimates all take on the expected signs (positive for employment and wages, negative for living in poverty). The estimated impact on employment is statistically significant at the 5 percent level and the other two estimates are just above and below the threshold of statistical significance at the 10 percent level. Following cohorts, the estimated effects reveal the predicted patterns as shown in the bottom of the table and the employment effects are statistically significant for the preschool cohorts in 1969.

31

The lack of any impact on educational attainment may be considered surprising in light of the positive effect on grade-for-age status that we identified earlier. However, whether a student falls behind in school at a young age does not have a very strong predictive relationship with whether he or she ultimately graduates high school. To formalize the strength (or lack thereof) of the relationship between grade-for-age status in 1980 with educational attainment in 1990, we collapsed data from the two Censuses to generate rates of each type of outcome by birth cohort and state and county of birth (noting our earlier discussion regarding geographic distinctions in the data). The correlation coefficient between rates of grade-for-age status in 1980 with the high school dropout rate in 1990 for the same birth cohort is -0.44. This is a substantial correlation, but also makes clear the possibility that Sesa me Street may have improved grade-for-age status without reducing the subsequent high school dropout rate for the same cohort.


The magnitude of these effects, though, is small.32 Our standard for interpreting magnitudes has been to evaluate the impact of moving from a weak reception county to a strong reception county, characterized by a 30 point increase in coverage. These results predict that employment would rise by about one percentage point (0.3*0.034). Similarly, wages would be predicted to rise by 0.93 percent (0.3*0.31; note that wages are measured in natural logs). Perhaps we should not expect to see large labor market effects driven by early childhood exposure to Sesame Street. To put these small estimated magnitudes into context, we consider what we might have expected to see given the magnitude of effects on early childhood test scores and grade-for-age status. To pursue this simulation exercise, we relate gains in test scores and grade-for-age status to wages, and then multiply this relationship by the estimated effects of

Sesame Street on these two measures. Previous estimates suggest that a 0.1 standard deviation increase in reading test scores is associated with a 0.6 percent wage increase (Levine and Zimmerman, 2010).33 The estimates from Bogatz and Ball (1971) indicate that Sesame Street increased reading test scores of viewers by 0.36 standard deviations. Our simulation is based on a 30 percentage point increase in exposure associated with greater reception. This would generate a prediction of a 0.65 percent (0.3*0.36*0.6) increase in wages. This is actually somewhat smaller than what we observe. 32

One potential concern with the longer term estimates for educational attainment and labor market outcomes as compared to the grade-for-age results is the possibility of greater attenuation bias in the 1990 and 2000 Censuses than in the 1980 Census. In the mobility results we reported in Table 2 using the NLSY, this would appear in the greater likelihood of a more distant inter-county move within states over time (attenuation bias in interstate mobility is not an issue based on our decision to examine just those who live in the same state as they were born). We used those data to simulate what would happen if 5.3 percent of the sample moved out of the area around their county of birth and we re-assigned them randomly to other locations. This amount is the difference in inter-county mobility observed between 1980 and 2000. The impact on our point estimates was minor, implying that greater attenuation bias cannot explain the stronger results in grade-for-age status than in educational or labor market outcomes. 33 Based on a review of past studies estimating the relationship between test scores and wages, Krueger (2003) FRQFOXGHV WKDW ³a plausible assumption is that a one SD increase in either math or reading scores in elementary schools is associated with about 8 percent higher earnings.´7KHHVWLPDWH ZHXVHLVVLPLODUWRWKLs, but it is likely more relevant since it is based on test scores at young ages in an American context (using the NLSY), which is not true of the other studies.


To consider what our estimated effects on grade-for-age status would imply, we need to document the relationship between this outcome and wages. Using data from the 1979 National Longitudinal Study of Youth (NLSY), we regress the natural log of wages in 2010 on grade-forage status in 1979 among those born in 1962-1964 (who should still be in school in 1979), controlling for race/ethnicity, gender, living arrangements, and parental education. The results indicate that students who are in the appropriate grade in high school earn 27 percent more than those who are behind. Our analysis of 1980 Census data, reported earlier, indicates that those in strong reception areas are 3 percentage points more likely to be in the appropriate grade relative to those in weak reception areas. This suggests we should find a 0.81 percent increase in wages associated with living in a strong reception area. Again, this is in the same vicinity as the wage effect that we estimated. What we can conclude from this analysis is that the results do not provide strong evidence of substantive improvements in labor market outcomes. This is not to say that exposure to Sesame Street had no effect; the fact that we are not able to strongly distinguish our estimates from zero may reflect limited power, even using large Census samples. V . A N A L YSIS O F H I G H S C H O O L

AND

B E Y OND D A T A

One limitation of our analysis so far is that it cannot capture specific mechanisms that have been explored more recently in thinking about longer-term effects of educational interventions (cf. Heckman, 2006). Interventions may be effective in improving narrower PHDVXUHV RI DFDGHPLFDFKLHYHPHQW³FRJQLWLYH RXWFRPHV´ and/or LPSURYLQJD FKLOG¶V level of socio-economic development (sometimes referred to by economists DV ³QRQ-cognitive RXWFRPHV´ Census data do not allow us to distinguish between these types of intermediate outcomes because no measures of non-cognitive outcomes are available.


To overcome this limitation, we augment the analysis above with an analysis of data available in the longitudinal survey, High School and Beyond (HSB). These data are obtained from a school-based survey of around 60,000 students who were high school sophomores and seniors in 1980. This data source contains extensive data on a range of outcomes including test scores, school grades, and self-reported measures of self-esteem and locust of control. In our analysis below, we have standardized test scores, measures of self-esteem, and locus of control so that each has a mean of zero and a standard deviation of one. The National Center for Educational Statistics (NCES) in the U.S. Department of Education does not provide access to any geographic identifiers with HSB data, not even with a restricted use agreement. However, a contextual file is available through ICPSR that provides details regarding county level unemployment rates, employment growth rates, and the like. We make use of those data to reverse engineer the HSB data to identify the county locations of the school. Data Appendix 2 describes the procedure we used. There are two cohorts of students in the HSB data who would have entered first grade right before and after the introduction of Sesame Street. High school seniors in the spring of 1980 who advanced on target through the educational system mainly would have been born in 1962; sophomores would have been born in 1964. This provides the possibility of within-school controls since sophomores, but not seniors, would have been exposed to Sesame Street before starting first grade. Comparing the two groups across areas that differ by Sesame Street coverage rates provides a plausible method of identifying the effects of the show. Despite the significant advantages of these data for our purposes, they do possess some limitations. First, as suggested by earlier discussions in this paper, a substantial minority of students do not advance through the educational system on pace with their birth cohort. Many


students will have fallen behind by tenth grade. This weakens the experimental design because those sophomores who had fallen behind may have been born before 1964 and, thus, would not have been exposed to Sesame Street before entering school. This amounts to a contamination of the treatment, introducing a bias towards zero in our analysis. A second limitation is generated because the school-based nature of the data omits those students who had dropped out already. Although dropout rates prior to 10th grade are reasonably rare, by 12th grade they are not. As such, sample selection plagues the control group in a way that is not replicated in the treatment group. To address this problem, we restrict our sophomore sample to those who begin their senior year two years later in 1982, making for a fair comparison between the remaining sophomores and seniors. It does, however, restrict the scope of our analysis to those who survive the educational system through the end of high school. We also find that very few students who make it to twelfth grade fail to graduate high school. Despite these limitations, the HSB data provide a useful opportunity to examine specific intermediate outcomes, and we proceed accordingly. The results of our analysis are reported in Table 7. The top panel focuses on measures of 12th grade academic achievement, including test scores in math, vocabulary, and reading, along ZLWKKLJKVFKRRO JUDGHVDQLQGLFDWRUIRUUHFHLYLQJPRVWO\$¶VDQG%¶VLQVFKRRO 7KHERWWRP panel focuses on measures of socio-emotional development, including locus of control, selfesteem, behavior problems, and work ethic PHDVXUHGLQHDFKFRKRUW¶VVHQLRU\HDU. We estimate IV models comparable in form to those described earlier using Census data, though we include school fixed effects, rather than county fixed effects, in this specification. The key explanatory variable is the interaction between the predicted FRYHUDJH UDWH LQ RQH¶V FRXQW\ ZLWK FRKRUW (sophomore or senior in the initial survey). We also include state by cohort fixed effects and a


series of other explanatory variables, which include basic demographics along with greater detail on socioeconomic status, like family income. The results from this analysis do not support the conclusion of an impact of watching

Sesame Street on academic of socio-emotional development by the end of high school for these students. Although some of the outcomes are estimated to improve for the treated group (sophomores) with higher coverage rates, these effects are generally not statistically significant (the effect on self-esteem is significant at the 10 percent level). This finding is not the result of low power. Consider, for instance, the estimated impact on math test scores, which has a standard error of 0.060. We would be able to reject the null hypothesis of no effect if the coefficient was around 0.12 at the 5 percent level. This coefficient would indicate that moving from a low coverage state to a high coverage state would increase math test scores of the exposed cohort (sophomores in 1980) by 0.036 of a standard deviation. One potential interpretation is that any effect of the show in either dimension had completely faded by the time one reached the latter stages of his or her high school career. V I. D ISC USSI O N This paper has documented the effects of exposure to Sesame Street programming content on indicators of early school performance, ultimate educational attainment, and labor market outcomes. Well-conducted studies at the time Sesame Street was initially introduced provided evidence that watching the show generated an immediate and sizeable increase in test scores. Building on this existing body of evidence, our analysis finds positive impacts on the educational performance of children who experienced their preschool years when Sesame Street was on television in areas with greater broadcast coverage. Specifically, such children achieved relative increases in grade-for-age status. This outcome largely represents improvements in


academic progress in elementary school, when students at that time were more likely to fall behind their appropriate grade level. This effect was particularly pronounced for boys and black, non-Hispanic children, along with those children who grew up in counties characterized by greater economic disadvantage. In that regard, Sesame Street satisfied its goal of preparing children for school entry, especially for black and disadvantaged children.34 Remarkably, the show accomplished that DWDFRVWRIDURXQGSHUFKLOGSHU\HDULQWRGD\¶VGROODUV The data do not provide conclusive evidence of longer-term effects on ultimate educational attainment or labor market outcomes. As described earlier, though, the small effects on wages are consistent with the magnitude of our estimated effect of Sesame Street on gradefor-age status, as well as earlier estimates of the impact of the show on early childhood test scores. What then could help boost that impact to produce substantive effects beyond elementary school? Although our analysis does not address this question, we offer two possibilities. From the academic side, interventions that are longer-lasting in nature may help. Again, exposure to

Sesame Street represents a one-time intervention with nothing intentional to follow. Students reach 1st grade better able to begin a reading curriculum, for instance, but at some point all students will have accomplished that goal. One interpretation of weak longer-term effects is that educational interventions need to be maintained and continuous so that students can build on the gains they have already attained to keep the momentum going.35

34

Recent research has also explored the ability of newer educational programming ( Martha Speaks, Super Why!, and others) developed by the Corporation for Public Broadcasting to improve test scores of preschool and kindergarten students in random assignment experiments (Linebarger, et al., 2009 and 2010). These analyses also find sizeable effects of viewing the shows. None of these interventions have been going on long enough to examine long-term outcomes. 35 Regarding MOOCs, an intervention like Khan Academy has that potential, for instance, to satisfy this concern, since it is not just a one-time intervention in a specific school grade.


Alternatively (or in addition), perhaps more emphasis needs to be concurrently placed on socio-emotional development and academic skills. Among the types of early childhood education interventions that were introduced during this period in the late 1960s (including Head Start and Perry Preschool), Sesame Street was perhaps the most narrowly targeted at the objective of improving educational achievement at its origin. This targeting is clear in the original stated objectives of the show (Cooney, 1967). Those objectives go beyond only an emphasis on academic achievement, incorporation of cultural development (arts and music) and awareness of basic emotions, but their broader focus on socio-emotional development is rather limited.36 Other interventions at that time and those introduced in the intervening years (including universal prekindergarten) contain far greater emphasis on both academic achievement and socio-emotional development. The role played by socio-emotional development is frequently emphasized in research regarding the impact of early childhood education, including that on Head Start, Perry Preschool, and Project Star (cf. Carneiro and Heckman, 2005, and Chetty, et al., 2011). Interventions that do not emphasize the importance of this form of development may be limited in their ability to generate long-term effects. If so, then perhaps a blended learning environment incorporating both electronic communication of educational content DQG WKH KXPDQ HOHPHQW WR DIIHFW WKH ³soft VNLOOV´ PD\ EH SUHIHUDEOH, and cost-effective. An understanding of the importance of these (or potentially other) alternatives is critical in designing future efforts to improve subsequent economic circumstances through early childhood education. As research and policy discussions continue to focus on early childhood education, we believe that the impact of Sesame Street deserves to be included along with Perry Preschool, Head Start and other programs. 36

Over time, Sesa me Street has broadened its focus to include greater emphasis on socio-emotional development (Davis, 2008), but those changes are not assessed here.


References

Akanegbu, Anuli (2013). ³Does the Khan AcDGHP\ 3DVV WKH 022& µ'XFN 7HVW¶"´ EdTech. http://www.edtechmagazine.com/higher/article/2013/04/does-khan-academy-pass-mooc-ducktest (accessed 2/9/2015). Bogatz, Gerry Ann and Samuel Ball (1971). The Second Year of Sesame Street: A Continuing Evaluation. Princeton, NJ: Educational Testing Service. %URZQ/HV ³µSesame Street¶'D]]OHVLQ7HOHYLVLRQ3UHPLHUH´ Variety. December 24. &DUQHLUR3HGURDQG-DPHV-+HFNPDQ ³Human Capital Policy´LQ-DPHV-+HFNPDQ and Alan B. Krueger (eds.), Inequality in America: What Role for Human Capital Policies? Cambridge, MA: MIT Press. &DVFLR (OL]DEHWK 8 ³School Progression and the Grade Distribution of Students: Evidence from the Current Population Survey´,=$'LVFXVVLRQ3DSHU Cascio, Elizabeth U. and Ethan G. Lewis (2006). ³6FKRROLQJDQGWKH$UPHG)RUFHV4XDOLI\LQJ Test: Evidence from School-Entry Laws.´ The Journal of Human Resources, 41(2): 294-318. Cascio, Elizabeth U. (2009). Do Investments in Universal Early Education Pay Off? Long-Term Effects of Introducing Kindergartens into Public Schools. NBER Working Paper 14951. &DVFLR(OL]DEHWK8DQG'LDQH:KLWPRUH6FKDQ]HQEDFK ³Expanding Preschool Access for Disadvantaged Children´LQ 0HOLVVD6.HDUQH\ DQG %HQMDPLQ + +DUULV HGV Policies to Address Poverty in America. Brookings Institution: Washington, DC. Chetty, Raj, John Friedman, Nathaniel Hilger, Emmanuel Saez, Diane Schanzenbach, and Danny Yagan (2011). ³+RZ'RHV